Allow non-default dataset configurations

[cg123]

Just a tiny change to make it possible to use HuggingFace datasets with multiple configurations. For example, to train on the 'enron_emails' configuration of The Pile, you could use this configuration:

datasets:
  - path: EleutherAI/pile
    name: enron_emails
    type: completion

[winglian]

🔥 This has been on my todo list, thank you! Does this breaks functionality if the dataset doesnt have a named configuration?

[cg123]

Nope! This should still work exactly the same as before if the config field is not present.

The default value of the name argument is None, so load_datasets should get the exact same arguments for existing configs. I've tested a partial run of the openllama 3b config to be sure.

[NanoCode012]

Hey, nice work! Could you also please add this somewhere in the Readme for visibility?

[NanoCode012]

Thank you for this example. Could you also add this name: parameter within the "all yaml options" below?